DOUBLY PENALIZED LOGISTIC REGRESSION FOR GENOMEWIDE ASSOCIATION STUDIES WITH LINEARLY STRUCTURED GENETIC NETWORKS

Abstract

This research aims to integrate linear structures of genetic networks into genomewide analysis studies (GWAS). Lasso penalized logistic regression is ideally suited for continuous model selection in case-control disease gene mapping, especially when the number of predictor variables far exceeds the number of observations. But it fails to consider the structure of genetic networks. Imposing an additional weighted fused lasso can further remove irrelevant predictors. Nesterov's method is employed to handle the high dimensionality and complexity of genetic data. It also resolves the non-differentiability problem of the lasso and fused lasso penalties. In simulation studies, this proposed method shows advantages in some cases compared with lasso and fused lasso. We apply this method to the coeliac data on chromosome 8.